1 Following up from 12/15/2022

Before we resume answering the question of “would bval of 2 low only be an improvement overall?”, let’s see if it would be helpful in ANY individual cases.

Quick summary of conclusions from this documet:

2 Prepare data

# Update: I realized that the cndx is assigned BEFORE wllq == 0 rows are removed
# Which affects the cndx assignment in some cases!!
# Just in cases there are any other nuances I am misssing, 
# I'm going to pull the cndx's directly from tcpl instead of using my home brew method

library(tcpl)
tcplConf(drvr = 'MySQL', user = Sys.getenv('INVITRODB_USER_RO'), pass = Sys.getenv('INVITRODB_PASS_RO'), host = Sys.getenv('INVITRODB_HOST_RO'), db = 'invitrodb')
tcplListFlds('mc1')
shafer.acid.tb <- tcplLoadAcid(fld = 'asid', val = 20)
shafer.acute.acids <- shafer.acid.tb[grepl('_acute_',acnm), acid]

# Get the raw activity values
mc0 <- tcplLoadData(lvl = 0L, fld = c('acid'), val = shafer.acute.acids, type = 'mc')

# Get the cndx's and merge
mc1 <- tcplLoadData(lvl = 1L, fld = c('acid'), val = shafer.acute.acids, type = 'mc')
nrow(mc0) == nrow(mc1) # TRUE
tcpl.dat <- merge(mc0, mc1, by = intersect(names(mc0), names(mc1)))

# Might as well pull the tcpl resp and bval while I'm here
# mc3 <- tcplLoadData(lvl = 3L, fld = c('acid'), val = shafer.acute.acids, type = 'mc')
# oh that's right... tcpll doesn't return the bval
library(RMySQL)
con <- dbConnect(drv = MySQL(), user = Sys.getenv('INVITRODB_USER_RO'), pass = Sys.getenv('INVITRODB_PASS_RO'), host = Sys.getenv('INVITRODB_HOST_RO'), dbname = 'invitrodb')
dbGetQuery(con, paste0('DESC mc3'))
mc3 <- dbGetQuery(con, paste0('SELECT * FROM mc3 WHERE acid IN(',paste0(shafer.acute.acids,collapse=","),');'))
dbDisconnect(con)
setDT(mc3)
mc3 <- merge(mc3, mc1[, unique(.SD), .SDcols = c('m1id','apid')], by = 'm1id', all.x = T) # add the apid
# is the bval the same for the up adn dn direction?
mc3[, .(length(unique(bval))), by = .(acid, apid)][V1 > 1] # empty -> so all are the same
bval.tb <- mc3[, unique(.SD), .SDcols = c('bval','apid','acid','pval')] # just going by acid instead of aeid since the bval is the same for the up and dn direction
tcpl.dat <- merge(tcpl.dat, bval.tb, by = intersect(names(tcpl.dat), names(bval.tb)), all.x = T)

# Can we add meta data now?
tcpl.dat <- tcplPrepOtpt(tcpl.dat)
# yay!
rm(list = c('mc0','mc1'))


# Calculate the alternative bvals
dat <- tcpl.dat
rm(tcpl.dat)
dat[wllt == 'n', .N, by = spid] # most DMSO, some water, no Media
dat[, bval.n2low := median(rval[wllq == 1 & (wllt == 'n' | (wllt == 't' & cndx %in% c(1,2)))], na.rm = T), by = .(acnm, apid)]
dat[, .N, by = signif(bval.n2low, 3) == signif(bval, 3)] # these shoudl be the same..
# cool, all are the same except for where NA
dat[is.na(bval)] # looks like a TON of wllq == 0 wells here
dat[, bval.2low := median(rval[wllq == 1 & (wllt == 't' & cndx %in% c(1,2))], na.rm = T), by = .(acnm, apid)]
dat[, bval.n := median(rval[wllq == 1 & wllt == 'n'], na.rm = T), by = .(acnm, apid)]

setkey(dat, NULL)
save(dat, file = 'L:/Lab/NHEERL_MEA/Carpenter_Amy/pre-process_mea_acute_for_tcpl/investigations/dat_consider_bval_2_low_concs_only_2022-01-04.RData')
load('L:/Lab/NHEERL_MEA/Carpenter_Amy/pre-process_mea_acute_for_tcpl/investigations/dat_consider_bval_2_low_concs_only_2022-01-04.RData')

# previous data source:
# load('lvl0_snapshots/mea_acute_lvl0_2020-07-29.RData')
# dat <-  mea_acute_lvl0

3 Check out a few individual apid

Kelly highlighted the apid where Tetrabutylammonium bromide was tested. Seemed like bval was abnormally low, and that using cndx 1 & 2 only might improve the bval. This is what prompted this entire investigation. So as a gut check, let’s see if using bval of cndx 1 & 2 only woudl be an improvement for this apid.

# Kelly highlighted a few cases where the bval is definitely low and is causing issues
# e.g. wherever Tetrabutylammonium bromide was tested
check.apids <- dat[chnm == 'Tetrabutylammonium bromide', unique(apid)]

dat[apid %in% check.apids, unique(.SD), .SDcols = c('acnm',grep('bval',names(dat),val = T))][order(bval.n2low)]
p <- ggplot(dat[apid %in% check.apids & !grepl('(LDH)|(AB)',acnm)], aes(x = bval.n2low, y = bval.2low, text = acnm)) +
  geom_point()+
  geom_abline(slope = 1, intercept = 0)+
  xlab('bval.n2low (current)')+
  ylab('bval.2low (proposed)')+
  ggtitle(paste0('Comparison of bval with vs without DMSO wells for all assay components\nfor apid ',check.apids,' (where Tetrabutylammonium bromide tested)'))
plotly::ggplotly(p, tooltip = 'text')
# Okay, so the bval of 2 low only would be a bit better for some endpoints,
# Particularly the weighted mean firing rate and the network burst number
# But others not really  helped.

# NOTE: I found that several cndx 1&2 wells from this apid have wllq == 0 and so had to be removed
# So for several endpoints, the bval of n+2low may not actually include many additional wells
# dat[apid %in% check.apids & cndx %in% c(1,2) & wllt == 't' & wllq == 1, .N, by = .(acnm)]
# approximately 28 per plate 
# There should be 6*2*3 = 36 per plate
# So significantly fewer than there could be, but still a decent number added to the pool

ggplot(dat[wllq == 1 & apid %in% check.apids & grepl('_burst_frequency_mean',acnm)], aes(x = cndx, y = rval)) +
  geom_jitter(aes(color = chnm), width = 0.2, height = 0)+
  ggtitle(paste0('Burst Frequency Mean rvals for ', check.apids))

dat[, is_edgewell := rowi %in% c(1,6) | coli %in% c(1,8)]
ggplot(dat[wllq == 1 & apid %in% check.apids & grepl('_burst_frequency_mean',acnm)], aes(x = cndx, y = rval)) +
  geom_jitter(aes(color = is_edgewell), width = 0.2, height = 0)+
  ggtitle(paste0('Burst Frequency Mean rvals for ', check.apids))

# WWHAAAT?? This is the EXCACT opposite of what I expect to see
# all of the edgwells are near 0, but the middle wells are the ones that are too low!!
# Perhaps this is from the days when the middle wells were over heating?

Takeaways:

The bval of 2 low only would be a bit better for some endpoints, Particularly the weighted mean firing rate and the network burst number But others not really helped.

Note: I found that several cndx 1&2 wells from this apid have wllq == 0 and so had to be removed

I am quite surprised that the edgewells seem to be doign worse than the middle wells for this apid!! Perhaps this is from the days when the middle wells were over heating?

Overall, conclusion is that the bval of cndx 1 & 2 only is not really helpful here

4 Are there ANY apid where bval of 2 only is a signficant improvement?

# Are there ANY apid where bval of 2 only is a signficant improvement?
all.bval.tb <- dat[, unique(.SD), .SDcols = c('bval.n2low','bval.2low','bval.n', 'apid', 'acid','acnm')]
ggplot(all.bval.tb[!grepl('(LDH)|(AB)',acnm)], aes(x = bval.n2low, y = bval.2low))+
  geom_point(pch = 1)+
  geom_vline(xintercept = 0)+
  geom_hline(yintercept = 0)+
  geom_abline(slope = 1, intercept = 0, color = 'blue')+
  ggtitle('Correlation of bval with vs without DMSO\nfor all MEA Acute apid-acnms')

# This honestly just looks like random noise
# I'm not seeing any trend or meaningful subset of points where the bval of 2 only is all that different
# I was particularly hoping to see some points where bval.n2low is a large negative number, and bval.2low helps raise it up
# but I'm not seeing a good deal of that
# In fact, there are 4 points where the bval.2low seems to make it even worse!

ggplot(all.bval.tb[!grepl('(LDH)|(AB)',acnm)], aes(x = bval.n2low, y = bval.2low))+
  geom_point(pch = 1)+
  geom_vline(xintercept = 0)+
  geom_hline(yintercept = 0)+
  geom_abline(slope = 1, intercept = 0, color = 'blue')+
  ylim(-100, 0)+
  xlim(-100, 0)+
  ggtitle('Correlation of bval with vs without DMSO\nfor all MEA Acute apid-acnms')
## Warning: Removed 1965 rows containing missing values (geom_point).

# Looking at the individual endpoint with the worst bval's
acnmi <- 'CCTE_Shafer_MEA_acute_cross_correlation_area'
ggplot(dat[acnm == acnmi], aes(x = bval.n2low, y = bval.2low)) +
  geom_point()+
  geom_abline(slope = 1, intercept = 0)+
  ggtitle(paste0(acnmi,'\nBval of 2 Lowest conc\'s only vs\nBval of n wells + 2 lowest conc\'s'))

# Still no real improvement


# So am I convinced? Is this enough evidence by which to... give up on trying bval.2low?

Doesn’t look like bval.2low would be of much help.

So am I convinced? Is this enough evidence by which to… give up on trying bval.2low?

5 What is the largest difference between these bvals? (or percent difference?)

# What is the largest difference between these bvals? (or percent difference?)
# in what direction? (closer or farther from 0? Currently above or below 0?)
dat[, bval_perc_diff := (bval.n2low - bval.2low)/abs(bval.n2low)]
# dat[, summary(bval_perc_diff)] # actually, we have so many bval's close to 0, this is not great
dat[, bval_diff := (bval.2low - bval.n2low)]
ggplot(dat[!grepl('(LDH)|(AB)',acnm)], aes(x = bval.n2low, y = bval_diff)) +
  geom_point(pch = 1)+
  geom_hline(yintercept = 0)+
  geom_vline(xintercept = 0)+
  ylab('bval.2low - bval.n2low')+
  ggtitle('Difference between current and proposed bval vs current bval\nfor all MEA Acute apid-acnms')

# Positive bval diff -> bval.2low is greater than bval.n2low
dat[bval.n2low < 0, summary(bval_diff)]
##     Min.  1st Qu.   Median     Mean  3rd Qu.     Max. 
## -48.6765  -0.8115   0.0000   0.2108   1.3446  60.4639
# max is an increase of 60
dat[bval.n2low < 0 & bval_diff >= 50, .N, by = .(apid, acnm, bval.n2low, bval.2low)]
# I wouldn't call this an improvment -> it's bringing the bval much further away from 0!

# Greatest amount by which bval.2low brings the bval closer to 0
dat[bval.n2low < 0 & abs(bval.2low - 0) < abs(bval.n2low - 0), max(abs(bval.n2low) - abs(bval.2low))]
## [1] 22.949
dat[bval.n2low < 0 & abs(bval.n2low) - abs(bval.2low) >= 22, .N, by = .(apid, acnm, bval.n2low, bval.2low)]
# okay, so this is the plate where it helps the most

dat[bval.n2low < 0 & abs(bval.n2low) - abs(bval.2low) >= 15, .N, by = .(apid, acnm, bval.n2low, bval.2low)]
# okay, so yay, i have identified a handful of plates where it really helps!!

Okay, so yay, i have identified a handful of plates where bval.2low would really help!!

However, I think I need to establish - how often would bval.2low bring the bval closer to 0 vs how often would it bring it farther from 0?

6 How often would using bval.2low instead of bval.n2low help vs hurt?

# Let's try to narrow in on the conclusion I am tryign to make below:
dat[, bval_2low_closer_to_0 := abs(bval.2low) < abs(bval.n2low)]
# dat[is.na(bval_2low_closer_to_0)] # ah, right, this is the plate with essentially no control wells
dat[, .(num_apid_acnm = length(unique(paste0(apid,acnm)))), by = .(bval_2low_closer_to_0)]
# well that's pretty straightforward!!

# But, I mostly only care about the cases where the bval.n2low is "really bad"
dat[, bval_n2low_cat := 'bval.n2low: -50 - 50']
dat[bval.n2low < -50, bval_n2low_cat := 'bval.n2low <-50']
dat[bval.n2low > 50, bval_n2low_cat := 'bval.n2low >50']
tb1 <- dat[!is.na(bval.n2low), .(num_apid_acnm = length(unique(paste0(apid,acnm)))), by = .(bval_n2low_cat, bval_2low_closer_to_0)]
dcast(tb1, bval_2low_closer_to_0 ~ bval_n2low_cat,  value.var = 'num_apid_acnm')

Review:

Conclusions:

One last thing: Would the bval of DMSO wells only actually be an improvement?

7 How does bval.n compare?

# Perhaps worth it to check out how bval.n compares as well? (I'm assuming it wouldn't be any better than )
ggplot(dat[!grepl('(LDH)|(AB)',acnm)], aes(x = bval.n2low, y = bval.n))+
  geom_point(pch = 1)+
  geom_abline(slope = 1, intercept = 0, col = 'blue')+
  ggtitle('Correlation of bval of with vs with the 2 lowest concentrations tested\nfor all MEA Acute apid-acnms')
## Warning: Removed 12384 rows containing missing values (geom_point).

# maybe I'm seeeing things... but is there slighlty more mass abvoe the line?

ggplot(dat[!grepl('(LDH)|(AB)',acnm)], aes(x = bval.n2low, y = bval.n))+
  geom_point(pch = 1)+
  geom_abline(slope = 1, intercept = 0, col = 'blue')+
  ylim(-100, 100)+
  xlim(-100, 100)+
  ggtitle('Correlation of bval of with vs with the 2 lowest concentrations tested\nfor all MEA Acute apid-acnms')
## Warning: Removed 42528 rows containing missing values (geom_point).

Maaaybe there is a tiny bit more mass above the line than below, but it’s only a marginal difference.

I would say this is not really an improvement (i.e., bringing the large negative bval’s closer to 0) -> maybe in a few cases, but only marginally so. Again, it would probably offer a significant improvement just as often as it would help

8 Next step:

Try out other normalizaiton methods (See outlined in previous html file ‘game plan’!)